[Question] What's the purpose of using 'dummy_tensor' in class EmbeddingOptimizer

[ZHANGWENTAI]

The self.dummy_tensor in class EmbeddingOptimizer is used in the method forward(), but the parameter is not explicitly used. After deleting this parameter, I found the training speed increase but the performance of model decrease. Further experiments show that the process of retrival embedding is turned into async. I wonder what is the purpose of passing the unused parameter self.dummy_tensor?